[llvm] [VPlan] Delay adding canonical IV increment. (PR #82270)

[via llvm-commits]

================
@@ -8902,29 +8913,26 @@ void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF,
   }
 }
 
-// Add the necessary canonical IV and branch recipes required to control the
-// loop.
-static void addCanonicalIVRecipes(VPlan &Plan, Type *IdxTy, bool HasNUW,
-                                  DebugLoc DL) {
+// Add the required canonical IV along with its loop branch, but w/o its
+// increment - which is introduced later.
+static void addCanonicalIV(VPlan &Plan, Type *IdxTy, DebugLoc DL) {
   Value *StartIdx = ConstantInt::get(IdxTy, 0);
   auto *StartV = Plan.getOrAddLiveIn(StartIdx);
 
   // Add a VPCanonicalIVPHIRecipe starting at 0 to the header.
+  // TODO: Introduce a separate scalar phi recipe that can be used for codegen,
+  // turning VPCanonicalIVPHIRecipe into an 'abstract' recipe which cannot be
+  // executed directly.
   auto *CanonicalIVPHI = new VPCanonicalIVPHIRecipe(StartV, DL);
   VPRegionBlock *TopRegion = Plan.getVectorLoopRegion();
   VPBasicBlock *Header = TopRegion->getEntryBasicBlock();
   Header->insert(CanonicalIVPHI, Header->begin());
 
-  VPBuilder Builder(TopRegion->getExitingBasicBlock());
-  // Add a VPInstruction to increment the scalar canonical IV by VF * UF.
-  auto *CanonicalIVIncrement = Builder.createOverflowingOp(
-      Instruction::Add, {CanonicalIVPHI, &Plan.getVFxUF()}, {HasNUW, false}, DL,
-      "index.next");
-  CanonicalIVPHI->addOperand(CanonicalIVIncrement);
-
   // Add the BranchOnCount VPInstruction to the latch.
+  VPBuilder Builder(TopRegion->getExitingBasicBlock());
+  // TODO: introduce branch-on-count during VPlan final (pre-codegen) lowering.
   Builder.createNaryOp(VPInstruction::BranchOnCount,
-                       {CanonicalIVIncrement, &Plan.getVectorTripCount()}, DL);
+                       {CanonicalIVPHI, &Plan.getVectorTripCount()}, DL);
----------------
ayalz wrote:

How about having BranchOnCount also be responsible for bumping the IV, at least initially? (Inspired by PowerPC's bdnz instruction; can call it BranchOnIncrementCount.) It could then feed back the canonical IV Phi across the back-edge, and possibly be split into a separate Add later to simplify code-gen.
Unclear if doing so helps achieve the desired simplification(?)

https://github.com/llvm/llvm-project/pull/82270